How Phonetic Change Happens

Josef Fruehwald

2014-10-08

Introduction

This Talk

  • How does a change in pronunciation propogate across a speech community?
  • When a pronunciation changes over time, what aspect of speakers’ knowledge of their language is changing?
  • What aspects of speakers’ cognition constrain a pronunciation change like this in the first place?

The Philadelphia Neighborhood Corpus

plot of chunk unnamed-chunk-1

The Philadelphia Neighborhood Corpus

  • Despite having such a long time domain, the PNC doesn’t quite directly observe the propogation of a pronunciation change across a speech community. We’ll still try to make some reasonable inferences on the basis of the data that we do have.

The Data

Speakers Transcribed (s) Vowels Measured
326 694,419 615,429
library(shiny)
runApp(system.file("appdir/means", package="phoneticChange"))

/ay/ Raising

General pattern

plot of chunk unnamed-chunk-4

What does it Mean

Population Shift?

plot of chunk population

What would it look like?

Population Shift

  • Speakers are not internally variable. You are either a raising speaker, or a non-raising speaker.
  • The speech community is highly variable.

What does it mean?

Lexicon

plot of chunk lexicon

What would it look like?

Lexicon

  • Speakers are internally variable. Sometimes they’ll say a raising word, and other times they’ll say a non-raising word.
  • The speech communuty isn’t very variable. Speakers should be more or less similar to eachother.

What does it mean?

Variable Rule

  • /ay/ = [+low]
  • with probability \(p\), ay \(\rightarrow\) [-low]/__[-voice]
    • beginning of the change, \(p=0\)
    • middle of the change, \(p=0.5\)
    • end of the change, \(p=1\)

What does it mean?

Variable Rule

plot of chunk phon_var

What would it look like?

Variable Rule

  • Speakers would be internally variable, alternatively raising and not raising /ay/ with some internally consistant probability.
  • The speech communuty wouldn’t be very variable. There would be some characteristic probability of raising for birth cohorts & social strata.

What does it mean?

Continuous Change

plot of chunk continuous

What would it look like?

Continuous Change

  • Speakers wouldn’t be very variable aside from the intrinsic variation involved in trying to hit a real-valued target.
  • The speech community wouldn’t very variable. There would be some characteristic pronunciation target for birth cohorts & social strata.

Totally No Stawmen Here

  • There are at least a handfull of examples of language change that did progress like each of the models described here.
  • But some theories of phonology, phonetics, and sound change do rule out the possibility of the final, continuous change model.
  • The first three models seem to be the ones most commonly assumed by new entrants to the field.
  • There is often covert variation between practitioners as to which model they’re assuming!

What it does look like.

Trend with speaker means

plot of chunk trend_with_means

Cohen’s D

plot of chunk cohen_plot

What it means

  • We can rule out the population shift model.
  • The lexicon shift model, and the variable rule models can’t be ruled out yet.
    • If speakers were sampling from two distributions, or two lexicons, their means would still shift continuously.

Distributional Properties

  • Both the kurtosis and the standard deviation of a mixture distribution should systematically co-vary with the mean.
library(shiny)
runApp(system.file("appdir/mixdists", package="phoneticChange"))

Distributional Properties

plot of chunk unnamed-chunk-9

Distributional Properties

plot of chunk unnamed-chunk-10

Distributional Properties

Simulation

plot of chunk unnamed-chunk-12

Distributional Properties

Simulation

plot of chunk unnamed-chunk-14

Distributional Properties

Simulation

plot of chunk unnamed-chunk-15

Distributional Properties

Simulation

plot of chunk unnamed-chunk-16

Distributional Properties

Comparsion to non-change

plot of chunk unnamed-chunk-17

Distributional Properties

Comparsion to non-change

plot of chunk unnamed-chunk-18

Distributional Properties

  • It doesn’t look like any of the categorical mixing models characterize the data well.
    • Looking just at the distributional properties of pre-voiceless /ay/, it doesn’t look like speakers are every sampling from two different distributions.
    • Comparing pre-voiceless /ay/ to the rest of /ay/, its distributional properties aren’t different, or at least not in the way expected.
  • The continuous shift model looks like it best characterizes the data.

Individuals or Communities?

Where is this change?

  • Is this change being mostly driven by individual speakers changing how they speak?
  • Or is it being driven by inter-generational shift?

Problem

  • The PNC has no longitudinal data for individual speakers.

Where is this change?

Problem

plot of chunk unnamed-chunk-21

Where is this change?

  ays_means %>% 
    filter(plt_vclass == "ay0") %>%
    gam(F1_n ~ ti(age) + ti(DOB) + ti(age, DOB), data = .) -> model

Where is this change?

plot of chunk unnamed-chunk-24

Where is this change?

plot of chunk unnamed-chunk-26

Where is this change?

plot of chunk unnamed-chunk-28

Where is this change?

  • The estimated age trends within birth cohorts seem relatively flat.
  • The change seems to mostly play out between generational cohorts.

Why?

Non-random errors in the signal

Ohala 1981

Ohala (1981)

Non-random errors in the signal

Pre-voiceless Shortening

  • Before voiceless consonants, vowels are shorter.
  • With less time to make such a big gesture for /ay/, speakers cut corners and pronounce the first part higher.

Offglide Peripheralizatiom

  • Before voiceless consonants, the [i] part of /ay/ is pronounced even higher and fronter than usual.

Non-random errors in the signal

A natural experiment

The difference between /t/ and /d/ is largely neutralized in certain contexts.

/t/ /d/
faithful write ride
flapping writer rider
/t/ /d/
rʌit raid
rʌiɾɚ raiɾɚ

Flapping Neutralization

plot of chunk unnamed-chunk-30

Flapping Neutralization

plot of chunk unnamed-chunk-31

Flapping Neutralization

plot of chunk unnamed-chunk-32

Flapping Neutralization

  • The non-random errors in the noise seem to be distributed differently between pre-faithful and pre-flapped contexts:
    • Duration: faithful and flapped /t/ > flapped /d/ > faithful /d/
    • Peripheralization: faithful /t/ > flapped /t/ and faithful and flapped /d/

The Change

plot of chunk unnamed-chunk-33

The Change

  • The degree of participation in the change appears not to be circumscribed by other continuous parameters of speech, by other, top-down cognitive factors.

Non-linear Bayesian Modelling

plot of chunk unnamed-chunk-34

Non-linear Bayesian Modelling

plot of chunk unnamed-chunk-35

Non-linear Bayesian Modelling

plot of chunk unnamed-chunk-36

Other Things I’ve Looked At

  • Word Frequency
  • Ratio with which a stem appears in a flapping vs a faithful context

Conclusions

Conclusions

  • It is possible for continuous properites of speech to change.
    • These continuous properties must be part of speakers’ knoweldge of their language.
  • These changes are not necessarilly circumscribed by other continuous properties of speech.
    • A top down, categorical aspect of speakers’ knowledge of their language appears to have played a crucial role at the initiation of this sound change.
  • There is still a lot of work to be done in understanding how and why changes like this happen, if our reasonable assumptions about them don’t seem to pan out.